Table 1:

Training times for the full (K = 5) versions of the MoDL architecture to achieve validation SSIM≥ 0.98. For stable training, MoDL with image space convolutions must be initialized using the weights learned for a K = 1 MoDL network. MoDL architectures trained with our joint layers require no pre-training. In total, using joint layers results in roughly an 8x speed-up over the pure image space approach.

MoDL Layer	Pre-Training (Hrs)	Training (Hrs)	Total (Hrs)
Image Convolution	19	12	31
Joint Layer	0	4	4